AI kitchen backsplash tile preview with a tile sample photo

A tile sample as a second reference

Describing a tile in words loses the pattern. Photograph the sample tile or save the supplier's image, then send it as a second reference next to your kitchen photo. The prompt says which image is the room and which is the tile.

Add a mask_url over the backsplash zone. Sume documents the mask on ChatGPT Image 2.5 only, and the same model accepts up to 16 references.

Sume Image API docs list ChatGPT Image 2.5 as openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. OpenAI's guide says to choose Sunburst where editing precision matters most and Flare for fast everyday generation, so these edits use Sunburst.

Ask for layout, not only pattern

Say how the tile lays out: subway brick pattern, herringbone, stacked. Give the grout color. Ask that the countertop edge, outlets, cabinets and window stay unchanged. Tell the model to scale the tile so one tile is about the size it is in real life, for example 7.5 by 15 cm.

import os
import requests

REFS = [
    "https://example.com/kitchen.jpg",
    "https://example.com/tile.jpg",
]
resp = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "openai/gpt-image-2.5-sunburst",
        "prompt": "Image 1 is a kitchen photo, image 2 is a tile sample. "
                  "Cover only the backsplash with image 2 in a brick layout "
                  "with light grey grout, tiles about 7.5 by 15 cm. Keep "
                  "countertop, outlets, cabinets and window unchanged.",
        "aspect_ratio": "auto",
        "mask_url": "https://example.com/backsplash-mask.png",
        "input_references": [
            {"type": "image_url", "image_url": {"url": u}} for u in REFS
        ],
    },
    timeout=60,
)
print(resp.status_code)
print(resp.json())

Layouts to compare

Use the same photo, mask and tile and change only the layout phrase.

Check scale and edges

Image models are not good at exact dimensions. Count tiles across the backsplash in the render and compare with your measured wall. Check the edge where tile meets countertop and the outlets, which are the first places a render drifts.

Inputs Sume checks before it spends anything

Reference and mask URLs must be public HTTPS; localhost, private-network and non-HTTPS URLs are rejected. Sume also checks every field against the model's catalog entry, so a field the model does not list returns 400 unsupported_parameter instead of being dropped without a word.

If you are unsure which fields a model accepts, GET /v1/images/models lists them, and GET /v1/images/models/{id}/endpoints returns the per-endpoint capabilities and pricing.

Cost, retries and slow calls

Each edit is one billed image when it completes, and nothing when it fails. Sume's docs say the amount in usage.cost is what the wallet is charged, with the 1.25 multiplier on provider list price already applied. That makes a retry cheap to reason about: a failed attempt costs zero.

Slow settings, such as 4K output, high quality or a large n, can push a call past the 30-second wait. Then the response is 202 with a job envelope rather than the image, and you fetch the result from the job endpoints.